The Face That Launched a Thousand Follows: Using AI to Predict Celebrity Potential

Journal of Marketing Research Scholarly Insights are produced in partnership with the AMA Doctoral Students SIG – a shared interest network for Marketing PhD students across the world.

What makes someone look like a celebrity even before they become famous? Conventional attractiveness may seem like the obvious answer, but not all spokespeople, influencers, and celebrities are highly attractive. People who do not conform to conventional beauty standards often achieve fame and success, suggesting that individuals perceive something beyond just attractiveness in celebrities and spokespeople. To describe this phenomenon, a recent Journal of Marketing Research study introduces Celebrity Visual Potential (CVP), which represents the probability that a face belongs to the celebrity rather than the non-celebrity category.

The researchers conducted five studies—a predictive deep learning model based on 22,000 face images, two controlled experiments, and two observational studies—to identify the facial features that predict CVP. Their model achieves 95.92% accuracy in classifying faces as celebrity or non-celebrity and produces a continuous CVP score for any face.

The findings suggest that facial width-to-height ratio, sexual dimorphism (masculinity for men, femininity for women), averageness, high cheekbones, and skin color are the five most important features predicting CVP. However, these features differ in the direction of their relationships with CVP: facial width-to-height ratio and averageness are negatively correlated with CVP, while sexual dimorphism, high cheekbones, and skin color are positively correlated. Notably, celebrities tend to have distinctive rather than average-looking faces. This finding contradicts the longstanding assumption that average faces are more attractive and, therefore, more “celebrity-like.”

Facial Features Signal Personality Traits

Beyond these physical features, the authors suggest that facial characteristics influence CVP by triggering personality trait inferences related to charisma, such as dominance, warmth, and competence. Two experiments using morphed faces confirm that model-predicted CVP aligns with human judgments of celebrity potential and that these personality-trait inferences (along with attractiveness) help explain why certain faces score higher on CVP. Crucially, CVP and attractiveness correlate at just .17 across more than 2,000 Instagram selfies, demonstrating that the two constructs capture fundamentally different qualities. Ultimately, the research shows that CVP is a comprehensive measure of the visual probability of gaining fame and influence, offering explanatory power beyond physical attractiveness.

Three key takeaways stand out for practitioners:

1. Attractiveness alone is an incomplete and idiosyncratic guide for selecting spokespeople, influencers, or brand ambassadors. CVP offers a richer and more predictive metric that reflects charisma-related trait inferences, not just physical appeal.
2. CVP predicts real-world outcomes. Instagram influencers with higher CVP experience greater follower growth, and C-suite executives at Fortune 500 companies have significantly higher CVP (.85) than non-C-suite employees (.20).
3. CVP is not tied to any demographic group—prediction accuracy does not vary significantly across race, gender, or age—suggesting it can be used as one input among many without reinforcing demographic biases.

However, the authors caution that CVP should never be the sole criterion in any selection decision, and organizations must obtain proper consent before collecting or using facial data.

We had the opportunity to correspond with Dr. Xiaohang (Flora) Feng, who responded on behalf of the author team, to learn more about their research and its practical implications.

Q: Your paper proposes that celebrity potential goes far beyond physical attractiveness, introducing the concept of Celebrity Visual Potential (CVP). How did the idea for CVP originate? Was there a specific moment or real-world observation that made you think attractiveness wasn’t the whole story?

A: The idea came from observing a simple paradox: many of the world’s most influential celebrities are not necessarily the most conventionally attractive, while many highly attractive people never become celebrities. That suggested to us that something beyond attractiveness was driving people’s perceptions of influence and charisma.

Marketing research has traditionally emphasized physical attractiveness as a key factor in spokesperson effectiveness, especially in low-involvement purchase settings. But psychology and sociology have long argued that charisma is multidimensional, involving traits such as dominance, warmth, competence, and trustworthiness. We wondered whether some of those impressions could be inferred from facial features before someone even speaks.

CVP emerged as an attempt to quantify that broader visual impression. Rather than asking, “Is this person attractive?”, we ask, “Does this face naturally evoke the kinds of personality inferences that make people believe this individual could become influential?” Our experiments showed that human judgments of celebrity potential closely tracked the model’s predictions, even after accounting for attractiveness. This suggests that people are indeed responding to something more complex than beauty alone.

Q: Part of the motivation for CVP is helping marketers select spokespeople across entertainment, media, and business contexts. How do you envision it being used in practice for that purpose, alongside brand fit and audience alignment?

A: I view CVP as an additional decision-support tool rather than a replacement for existing marketing criteria.

Choosing a spokesperson has always required balancing many factors: brand fit, authenticity, expertise, audience demographics, reputation, and campaign objectives. Our results suggest that CVP adds incremental information by capturing first impressions related to perceived charisma and persuasive potential.

For example, if two potential influencers have similar audience reach and brand alignment, CVP may provide another signal about which individual is more likely to capture attention or generate engagement. Our observational studies found that CVP predicted future follower growth on Instagram and distinguished C-suite executives from other employees on LinkedIn, even after controlling for attractiveness and other contextual variables.

That said, the paper also emphasizes that CVP explains only a modest portion of performance. Marketing success depends on many factors beyond appearance, including creativity, communication skills, authenticity, and long-term relationships with audiences. CVP should therefore complement, not replace, human judgment.

Q: This project spans deep learning, face morphing, large-scale image datasets, controlled experiments, and observational validation across two platforms. What was the most unexpected challenge you faced during the research process, and is there anything you’d do differently if you were starting today?

A: The biggest challenge was balancing predictive accuracy with interpretability.

Modern deep learning models can achieve impressive predictive performance, but they are often criticized as black boxes. We didn’t want to publish a model that simply predicted celebrity status without explaining why. That motivated us to develop the FeatureSHAP framework, which allowed us to estimate the contributions of different facial characteristics to CVP while retaining the predictive power of deep learning.

Another challenge was validating that the model reflected human perception rather than simply memorizing celebrity faces. That’s why we conducted controlled experiments using face morphing, in which participants evaluated entirely new faces with systematically manipulated CVP levels. The fact that human judgments increased almost linearly with model-predicted CVP gave us confidence that the model was capturing meaningful psychological signals rather than just recognizing famous identities.

If I were starting today, I would probably build a much larger and more globally representative dataset. Beauty standards, media exposure, and celebrity cultures differ across countries, so expanding the diversity of both the training data and validation settings would make the model even more broadly applicable.

Q: Your study finds that sexual dimorphism (masculinity for men, femininity for women) positively predicts CVP. Yet beauty standards are shifting: androgynous aesthetics are increasingly prominent in high fashion and popular culture. At a societal level, do you foresee CVP needing periodic recalibration as cultural ideals evolve, or do you think the underlying mechanisms are stable even as surface-level beauty trends shift?

A: I think recalibration is both necessary and scientifically healthy.

One important point is that our model was trained on historical data reflecting how celebrities have emerged in existing media environments. Like any machine learning model, it learns from observed patterns rather than prescribing what society should value.

Some underlying psychological mechanisms, such as the rapid formation of impressions of competence, warmth, or dominance from faces, appear relatively stable across decades of psychological research. Many of these findings have been explained from an evolutionary perspective, suggesting that humans may have developed certain facial heuristics because they were adaptive for making quick social judgments about allies, leaders, or potential threats. If these mechanisms are indeed rooted in evolutionary psychology, they are likely to be more stable than short-term cultural trends.

At the same time, the way those first impressions translate into celebrity status is shaped by culture. Media platforms change, audiences change, and the types of public figures who become influential also evolve. The increasing popularity of androgynous aesthetics is a good example. Even if people’s underlying tendency to make rapid personality inferences remains relatively stable, the cultural meaning attached to particular facial characteristics can shift over time.

I therefore view CVP as a dynamic measure rather than a timeless definition of celebrity. Periodic retraining using contemporary data is an important part of responsible deployment as it allows the model to adapt to evolving cultural ideals while still capturing the more fundamental psychological mechanisms underlying first impressions.

Q: How do you envision organizations responsibly deploying CVP in practice—for example, in hiring or influencer selection—and what safeguards would you recommend to prevent misuse or reinforcement of appearance-based biases? Are there contexts where you would actively discourage its use?

A: Responsible deployment is extremely important, and we explicitly discuss this in the paper.

First, CVP should only be considered in contexts where visual persuasion is genuinely part of the role, such as advertising, influencer marketing, entertainment, or the selection of public-facing spokespeople. Even in those cases, it should be only one input among many.

Second, organizations should never use CVP as an automatic screening or hiring criterion. Professional competence, experience, creativity, communication ability, and ethical judgment remain far more important predictors of success in most occupations.

Third, transparency and consent matter. Facial data are highly sensitive, and organizations should obtain explicit permission before analyzing images.

Finally, there are contexts where I would actively discourage its use. Jobs that have little to do with public persuasion, such as engineering, accounting, scientific research, or many administrative positions, should not incorporate appearance-based metrics. Likewise, I would oppose using CVP as the sole basis for employment decisions because it risks reinforcing existing appearance-based inequalities rather than helping organizations make better marketing decisions.

Q: If you were teaching this paper in a marketing class, what’s the biggest misconception you’d want to correct: about celebrity, about AI, or about the role of appearance in marketing?

A: I would probably correct two misconceptions simultaneously.

The first is that celebrity is simply another word for attractiveness. Our findings suggest that attractiveness matters, but it explains only part of the story. People also infer competence, warmth, dominance, and other characteristics almost instantaneously from faces, and those combined impressions help shape perceived celebrity potential.

The second misconception is that AI discovers objective truth. AI learns patterns from historical data. If society changes, or if the training data are biased, the model will reflect those changes or biases. Therefore, AI should be viewed as a measurement tool, not as an authority that defines who deserves success.

Ultimately, I hope students take away a broader lesson: understanding human perception is important in marketing. AI can help us measure those perceptions more systematically, but it should always be combined with theory, experimentation, and thoughtful human judgment.

Source: Xiaohang Feng, Shunyuan Zhang, Xiao Liu, Kannan Srinivasan, and Cait Lamberton (2025), “An AI Method to Score Celebrity Visual Potential,” Journal of Marketing Research, 62 (5), 757−75. doi:10.1177/00222437251323238

Go to the Journal of Marketing Research